From Learned Perception to Control
A detector's output is not yet a useful motion command. Give its coordinates a frame and a timestamp, then follow Botie through repeated corrections. When the actuator cannot do what the controller asks, repair the controller's memory as well as its output.
Making predictions usable for motion
The camera’s position reading refers to a particular view and time. We need to account for both before using it to move towards an object.
- 01Ready to exploreThe Detector Found the Object. Where Is It for the Robot?
Transform a learned location from camera coordinates into the robot’s coordinate frame.
- 02Ready to exploreA Correct Detection Can Already Be Out of Date
Propagate a timestamped position estimate under an explicit motion assumption and reject stale observations.
Closing the action loop
An action changes what the next observation will show. Use that feedback to adjust the next command, then see what happens when the motor cannot apply the command we requested.
- 03Ready to exploreMeasure Again Before Choosing the Next Move
Use feedback to correct position errors and inspect how gain and timestep affect a simple closed loop.
- 04Ready to exploreRemember What the Actuator Could Actually Do
Add integral correction, then handle saturation with an explicit anti-windup update.